Re: fixed location of share info page in HVM guests

[Keir Fraser <keir.xen@gmail.com>]

On 27/08/2012 22:56, "Olaf Hering" <olaf@aepfle.de> wrote:

> On Mon, Aug 27, Keir Fraser wrote:
> 
>> On 27/08/2012 22:32, "Olaf Hering" <olaf@aepfle.de> wrote:
>> 
>>>>> Recently I tried to move the shared_info page in the pvops kernel during
>>>>> shutdown, see "xen PVonHVM: move shared_info to MMIO before kexec"
>>>>> patch. But this change was reverted because it caused reboot failures
>>>>> because the actual moving of the shared page is fragile.
>>>> 
>>>> How was it fragile? Moving it into MMIO should just work?
>>> 
>>> The moving itself is not the issue, but the possible access to the page
>>> during the move. Its not atomic, nor can it easily be atomic.
>> 
>> Why not map it into MMIO in the first place, rather than into the middle of
>> RAM? Do that early during boot, and early during resume from
>> save/restore/migrate (i.e., places you presumably already map the
>> shared_info page, but into the middle of RAM)?
> 
> I think there is no easy way to know where the PCI device is located at
> the time the shared info page is configured in the early kernel setup.

How about we guarantee that guests can use the 1MB in address range
0xFED00000-0xFEDFFFFF for whatever mappings they like, guaranteed unused (or
at least, mapped with nothing useful) when guest kernel starts?

This is already, and has always been, the case. But we can etch it in stone
quite easily. :)

Then you can stick shared_info at fed00000 always, plus there's space there
for plenty of other stuff that might one day be useful to map very early.

 -- Keir

> Olaf